[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[DooKey]

Unless they're holding back something big, this is a colossal disappoint and a terrible release.

How could I possibly be excited about a gpu about as fast as the GTX 1080 from May 2016 (more than 3 years earlier than its launch date) for $100 less than a GTX 1080 was in May of 2016. If this is all they have to show, its a total failure.

64 CU - AGAIN. What is wrong with their design team??

Just imagine if NV had spent their whole die budget on rasterization instead of using some for RT. I bet leather jacket is pissed at himself. He could have buried AMD this generation.

 

[railven]

Two things:

1) die size much larger than I was expecting.

2) RDNA? Really? Ok.

 

[Glo.]

There is no indication that it is a clean sheet design. AMD might just wanted to walk away from GCN naming for its gaming GPUs since its reputation isn't so great in the first place.

Once again. I've posted it time, and time and again:

GCN is redactedINSTRUCTION SET. Its like saying x86 has bad reputation. Architecture was named GCN, because it was designed after GCN ISA!

New CUs, new front end and their words that it was cleen sheet like Zen 1 design indicates that was indeed cleen sheet design.

Aside from the fact, that making implications for a new uarch are difficult indeed, we still can safely assume, that AMD have chosen a game, which performs way above average compared to the competition for marketing purpose. Since the demonstrated margin of 10% is not really large, it is at this point almost save to assume that RX5700 is slower than RTX2070 in average.
This comes on top of complete lack of raytracing acceleration - for all we know DXR is not supported - not even as SW/GPGPU driver.

How is it safe to assume that it is slower than RTX 2070? based on what? We do not have any clue how it will behave with its new front end. How new CUs will behave. We even do not know wheter RX 5700 is the highest SKU in the lineup.

 

[Mopetar]

Just imagine if NV had spent their whole die budget on rasterization instead of using some for RT. I bet leather jacket is pissed at himself. He could have buried AMD this generation.

It's best to introduce new technology like RT when the competition is weak. If AMD were competitive, they wouldn't be able to spare the silicon since they'd need to devote more to keeping pace with AMD. It's a bit of a gambit since NVidia still needs to refine the technology and hope that AMD doesn't meet or surpass them in terms of conventional performance, otherwise they've got a lot of dedicated hardware the adds to the cost that most customers won't care about.

 

[Ajay]

Ain't that the truth. Everyone who was in the market for Polaris over the last three years is saying the saaaaame thing.

Can we get Navi at $279 or less, please? Preferably $239 or less?

Why do people want AMD to lose money? If the performance is really @ a RTX 2070 level, give or take - then so should the price ($500-550). Maybe $50 less due to inferior brand recognition.

 

[Thala]

How is it safe to assume that it is slower than RTX 2070? based on what? We do not have any clue how it will behave with its new front end. How new CUs will behave. We even do not know wheter RX 5700 is the highest SKU in the lineup.

As i tried to explain, it is a pure statistical argument. It reasons about how much the mean is below the best case, given that best case is known. Thing is, that AMD did not chose a random sample but a good sample.

 

[SPBHM]

Why do people want AMD to lose money? If the performance is really @ a RTX 2070 level, give or take - then so should the price ($500-550). Maybe $50 less due to inferior brand recognition.

the RTX 2070 is not very good value, an AMD card of similar performance will have to face the 1660 Ti and 2060 better value just like the 2070, and the AMD card doesn't support RT like the 2060/2070

about losing money... if they price it at $499 and it's like a 2070 without RT and Nvidia branding it's not going to sell, a better price will give them volume making money for them.

 

[Thala]

Just imagine if NV had spent their whole die budget on rasterization instead of using some for RT. I bet leather jacket is pissed at himself. He could have buried AMD this generation.

With this sentiment you would never introduce revolutionary technology due to low adoption rate in the beginning. Rasterization was the method of choice since the beginning of HW accelerated gfx, because it was algorithmically cheap and not because it was the best known method.
Today we are at a point, where higher resolutions have diminishing returns - so the industry needs to reevaluate the methods going forward.

 

[Glo.]

As i tried to explain, it is a pure statistical argument. It reasons about how much the mean is below the best case, given that best case is known. Thing is, that AMD did not chose a random sample but a good sample.

What statistics you have to back that up, considering, there is no Navi benchmarks/reviews, comparing it to previous generations of GCN?

 

[Ajay]

the RTX 2070 is not very good value, an AMD card of similar performance will have to face the 1660 Ti and 2060 better value just like the 2070, and the AMD card doesn't support RT like the 2060/2070

about losing money... if they price it at $499 and it's like a 2070 without RT and Nvidia branding it's not going to sell, a better price will give them volume making money for them.

If Navi has 2070 performance, why would it have to compete with the 2060, or, even more bizzarely, the 1660 Ti??! I'm sure Navi will have a cut down (die scavenged) lower CU count card to go up against NVs other offerings. RT, at this point, is mostly marketing buzz - there are literally thousands of games that don't support RT.

I think Navi will sell against NV's line up unless the performance just doesn't match up when it comes to benchmark comparisons. AMD desperately needs to get higher margins, and I wish them the best. Maybe Intel will be cheaper.

 

[SPBHM]

If Navi has 2070 performance, why would it have to compete with the 2060, or, even more bizzarely, the 1660 Ti??! I'm sure Navi will have a cut down (die scavenged) lower CU count card to go up against NVs other offerings. RT, at this point, is mostly marketing buzz - there are literally thousands of games that don't support RT.

I think Navi will sell against NV's line up unless the performance just doesn't match up when it comes to benchmark comparisons. AMD desperately needs to get higher margins, and I wish them the best. Maybe Intel will be cheaper.

the 2070 also competes with the cheaper cards from Nvidia,
I assume AMD buyers on average are also more value oriented, the 2070 is not good value, an AMD version of that is going to be even worse (no RT)

so yes, someone buying a 2070 might look at the 2060 and notice that it's not that far in performance and a lot cheaper, same will happen to Navi if it's not priced correctly... and these cards are old news,

 

[Glo.]

Again, if "medium" Navi is RTX 2070+10/15%, then small Navi, being 50% of it between GTX 1660 and 1660 Ti. Small Navi can perfectly compete with the mainstream GTX GPUs from Nvidia.

 

[JDG1980]

I actually don't think that is a bad prediction, but I think those are best case scenarios. I think it'll be 175-190w TDP and be AT MOST 10% faster than Vega64.

For a new architecture *plus* a new node, that kind of meager gain over Vega 10 would be truly pathetic. But from RTG it would not necessarily be unexpected.

 

[JDG1980]

They launched VEGA Vii at the same MSRP as RTX2080, now that they will have a more competitive card (both in performance and power consumption) against a $499 MSRP RTX2070 they will price it at $299 ??? No way, $399 and $499 are still low volume segment, they dont want to handicap NAVI 14 that will be way smaller and compete directly against RTX2060 by lowering the MSRP of the low volume RTX2070 competitor.

If past precedent is anything to go by, they've got a card that can *marginally* edge out GTX 1070 in the AMD-friendly titles, but will almost certainly fall behind it, perhaps substantially, in the Nvidia-friendly titles. In addition to that it will lack Nvidia's special features (RT/DLSS) - and while those may not be the world's most popular, they are certainly worth *something*. Performance-per-watt might be roughly competitive this time, just because RTG will have a full node advantage - or maybe they'll sacrifice that again and blast it to 225W just to eke out another percentage point or two. And OpenGL support on Windows will suck again, so forget about emulators like CEMU.

If AMD releases that at $399, people will yawn. It has to be $329 at most.

 

[JDG1980]

There is no indication that it is a clean sheet design. AMD might just wanted to walk away from GCN naming for its gaming GPUs since its reputation isn't so great in the first place.

Lisa Su said during the presentation that it was a clean sheet design.

 

[JasonLD]

Lisa Su said during the presentation that it was a clean sheet design.

Actually, no, I don't think she ever mentioned it of being a clean sheet design. There is a reason why Anandtech article is still on a fence about whether it is major overhauled architecture or not. Well, we will find out in a month or two.

 

[SPBHM]

Actually, no, I don't think she ever mentioned it of being a clean sheet design. There is a reason why Anandtech article is still on a fence about whether it is major overhauled architecture or not. Well, we will find out in a month or two.

she heavily implied at least, she said Zen was a "from scratch" design followed by " that's exactly what RDNA is"

 

[tviceman]

For a new architecture *plus* a new node, that kind of meager gain over Vega 10 would be truly pathetic. But from RTG it would not necessarily be unexpected.

Stop comparing it to Vega 10. Compare it to the Polaris RX 480, before AMD threw caution out to the wind and factory overclocked all their cards to the max with the RX 580 or the respun 590. If they literally doubled the performance of the RX 480, it would be about 6% faster than Vega 64. Doubling of performance would be pretty damn impressive. Nvidia didn't come close to that with either Kepler from Fermi (new arch and new node) or with Pascal from Maxwell (also new arch with new node).

 

[Thala]

What statistics you have to back that up, considering, there is no Navi benchmarks/reviews, comparing it to previous generations of GCN?

Where in my argument does it play any role how Navi compares to GCN? I have one sample point which shows RX5700 10% ahead of RTX2070 - then i reason about peak-to-average ratios being larger than 10% in essentially any generation before (or even when comparing 2 random GPUs from any generation) and finally concluding that RX5700 is slower than RTX2070 in average is likely.

To counter this statistical argument, show me one example from history of GPUs where peak-to-average performance ratio is less than 10% between 2 GPUs.

 

[JDG1980]

Stop comparing it to Vega 10. Compare it to the Polaris RX 480, before AMD threw caution out to the wind and factory overclocked all their cards to the max with the RX 580 or the respun 590. If they literally doubled the performance of the RX 480, it would be about 6% faster than Vega 64. Doubling of performance would be pretty damn impressive. Nvidia didn't come close to that with either Kepler from Fermi (new arch and new node) or with Pascal from Maxwell (also new arch with new node).

How acceptable the performance is depends on three factors: die size, power consumption, and price.

Polaris 10 was the successor to Pitcairn - both were low/midrange chips with a similar die size (212mm^2 for Pitcairn, 232mm^2 for Polaris 10) and a 256-bit bus width. Performance improvements from R9 270X (Pitcairn) to RX 480 (Polaris 10) ranged from 85% at 1080p to 96% at 4K (source). And the R9 270X is more power-hungry, maxing out around 175W when the RX 480 barely breaks 150W. This indicates that a full process node gain plus improved architecture should be able to nearly double performance, at least given RTG's modest starting point. RTX 2070 is about double RX 480's performance, so this is a sensible target, and not just because AMD put it up against RX 5700 in the keynote.

drchoi71 on Reddit put together a photo of the leaked PCB with the new Navi die at correct scale. From that, he estimated about 252mm^2 for Navi's die size. If that is accurate, then it strengthens the argument to treat it as the 7nm successor to 14nm Polaris 10 and 28nm Pitcairn. We are pretty sure from the PCB that it is going to have a 256-bit memory bus, this time with GDDR6. Let's give AMD the benefit of the doubt (usually OK on the CPU side these days, but a risky thing on the RTG side) and assume that if it beats RTX 2070 by 10% on an AMD-favored benchmark, it can at least come within 5% of RTX 2070 in overall gaming performance averaged over a variety of titles (granting that a few games on the other side will favor Nvidia heavily instead). That leaves power consumption and price as the wildcards.

When Pitcairn debuted, the top SKU (7870 GHz Edition) cost $349 and maxed out at 144W of power consumption (though it was technically rated for 175W TDP). When Polaris 10 debuted, the top SKU (RX 480 8GB) cost $239 and maxed out at 167W, violating its formal 150W TDP (I believe a firmware update later fixed this).

Therefore, we have a right to expect that the top RX 5700 SKU should come to the table at a price point of $239-$349, and a TDP around 150W. If they blow up the power budget way beyond that, or try to price it at $399 or above, I think that would count as a failure from a consumer perspective.

 

[Mockingbird]

Agreed. I think the leak prices of $399 for Navi14 and $499 for Navi10 are both about $100 too high, though lowering them both by $50 wouldn't be a bad idea either.

Somewhere along the line, they were supposed to replace Polaris. RX480 launched at $239 for the 8GB version. RX470, $179. RX580, $229 for 8GB version. RX570, $169. RX590 was the most expensive Polaris at $279.

Where is the Polaris replacement?

Said who?

Navi is supposed to replace Vega.

AMD already stopped production of Vega, with the intention of replacing it with Navi.

By contrast, AMD hasn't stopped production of Polaris at all.

 

[DrMrLordX]

I think anyone expecting AMD to come in cheap on prices is going to be disappointed. We know what they're selling the Ryzen 3000 CPUs for, and while you could argue that could be a result of supply constraints, it does tell us that if AMD has a competitive product, it will be priced that way.

The Ryzen 3xxx series is well-priced, using prices from 2017 almost exactly. R5 3600 is quite a deal, especially for anyone who overclocks it.

Why do people want AMD to lose money? If the performance is really @ a RTX 2070 level, give or take - then so should the price ($500-550). Maybe $50 less due to inferior brand recognition.

Navi was never supposed to be "at the RTX 2070 level". They needed to bury the 2060 and 1660Ti. The 1660Ti in particular looks like one of NV's highest-volume cards for the next half-year. Something that edges out the 2060 for $279 - which is the price lowpoint for the 1660Ti right now - would still give AMD some decent volume and margin. AMD would not lose money on such a card. They could have built the card with GDDR5 or GDDR5X to save money, and if the die were small enough, the 7nm process wouldn't matter cost-wise so much. The low-power nature of the card means no overbuilt VRMs, etc. It would not be that expensive of a card!

That way they gain money AND marketshare, and they throw off NV who can only respond with Ampere.

Instead we get a card which beats the 2060 for, what, $399? And by how much? And a card that matches (but apparently doesn't beat) the 2070 for $499? Who would pay such prices? Both the 2060 and 2070 are overpriced. Why would AMD target their price points? They don't have the tech or the brand recognition to push such high ASPs. At least Radeon VII was a one-off halo product with a hobbyist element (1/4 FP64) so it was not that big of a deal. Navi was supposed to be their bread and butter. Go for the high-volume segment, AMD! Not the 2070. As it stands, Navi10 for $499 and Navi14 for $399 will lose AMD money because they will not sell enough units.

Said who?

Navi is supposed to replace Vega.

AMD already stopped production of Vega, with the intention of replacing it with Navi.

By contrast, AMD hasn't stopped production of Polaris at all.

Polaris is in need of retirement. It is older than RX Vega. I get that they needed to stop production of RX Vega because of HBM. Polaris needs to go away, and Navi was supposed to put it out to pasture.

Now we have the RX590 anchoring AMD's midrange lineup for who knows how long. Pathetic.

 

[crazzy.heartz]

Something like a 48CU/64Rop/8 SE die with gddr6 and lots of small improvements all round would be preferable imo.

Spot on. This is at least a 256mm^2 chip, if not 275mm^2, as predicted by Ryan;

which makes it 25% smaller than the other 7mm AMD Chip (Vega 20)
@ 25% less than Vega 20 (331mm^2/64CUs) = ~ 250mm^2 / 48CUs

However, @ 275mm^2 or higher, it would be a 56 CU chip
@ 15% less than Vega 20 (331mm^2/64CUs) = ~ 280mm^2 / 56CUs

Also, if this is a 56CU chip and the yields are not good, then AMD can cut it thrice; in 8CU increments (56 / 48 / 40 CU parts)

They can launch the 40 & 48 CU chips now and build up an inventory of full dies. Launch these cut down parts now, where 40 CU competes with Rx2060 and 48CU part competes with Rx2070.

Then, once they've built up a healthy inventory & Nvidia launches a Rx2070Ti, they can simply release the full 56CU chip.

 

[railven]

So, before this thing goes any further, do we know if it's a clean sheet design or not? Been seeing a lot of "can't compare it to anything before, it's 'new'" and now a few "who said it's new?"

EDIT: Also, god reading RDNA now in other places hurts. What a terrible name. Welps, RDNA is here to stay. Hope it lives a better life than GCN did.

 

[DrMrLordX]

So, before this thing goes any further, do we know if it's a clean sheet design or not? Been seeing a lot of "can't compare it to anything before, it's 'new'" and now a few "who said it's new?"

Answer hazy, ask again later.

We don't really know. Maybe it is clean sheet, maybe it is not. The general assumption of most people following RTG is that Navi would be the last GCN-related design before they went clean sheet in 2021 (or whenever "generation next" comes out). Navi might be a stepping-stone to that new design. Is anything really clean sheet anymore?